Power distribution network edge gateway calculation decision-making system based on digital intelligent multilevel model

Abstract

The invention provides a power distribution network edge gateway calculation decision-making system based on a digital intelligent multilevel model, and relates to the technical field of power system intellectualization, and the system comprises a perception control layer which is used for collecting perception data of a low-voltage power distribution network; the edge intelligence layer comprises multiple stages of models which are connected in sequence, and the multiple stages of models comprise a data processing platform, a digital intelligence model library, a cross validation layer and a collaborative decision-making layer; and the cloud platform is used for full-amount data storage, global collaborative scheduling and visual display, and carrying out bidirectional parameter interaction with the edge intelligent layer. The edge decision-making capability of the novel power distribution network can be remarkably improved, the cloud computing load is reduced, and a powerful guarantee is provided for safe operation of a power system.

Description

Technical Field

[0001] The present invention relates to the field of intelligent power system technology, and in particular to a distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model. Background Art

[0002] Against the backdrop of large-scale integration of new energy sources and increasing power system complexity, traditional distribution networks face a series of severe challenges. For example, they suffer from significant shortcomings in real-time performance, highly concentrated computing pressure, and a lack of effective security mechanisms for data transmission and storage on edge devices.

[0003] Currently, existing technologies, including centralized systems, independent edge systems, and conventional cloud-edge collaborative architectures, all exhibit certain shortcomings. Centralized systems rely heavily on global optimization in the cloud, but this lacks real-time performance, imposes high bandwidth costs, and presents significant challenges in processing heterogeneous data at the edge. While independent edge systems can achieve real-time responses, their lack of global coordination makes decisions prone to local optimality rather than achieving a global optimum. Conventional cloud-edge collaborative architectures are relatively simple in structure and lack intelligent processing capabilities, making it difficult to compensate for the decline in prediction accuracy caused by external environmental factors in real time.

[0004] In view of this, developing a decision-making system that is adapted to edge gateways and has both lightweight and intelligent characteristics is of great practical significance. Summary of the Invention

[0005] To this end, the present invention provides a distribution network edge gateway computing decision-making system based on a digital intelligence multi-level model, which is used to solve the problems in the existing technology of insufficient real-time performance of traditional distribution networks, concentrated computing pressure, lack of edge device data transmission and storage security mechanism, as well as the existing centralized systems, edge independent systems, and cloud-edge collaborative architectures, which have poor real-time performance, lack of global collaboration, and insufficient intelligent processing.

[0006] To solve the above problems, an embodiment of the present invention provides a distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model, which includes: The perception control layer is used to collect perception data of the low-voltage distribution network, including operation data, equipment status data and external environment data; The edge intelligence layer includes a multi-level model connected in sequence. The multi-level model includes a data processing platform, a digital intelligence model library, a cross-validation layer, and a collaborative decision-making layer, where: The data processing platform is used to pre-process the perception data, including error handling, anomaly detection, data conversion and feature extraction, data fusion and protocol encapsulation; A digital intelligence model library, used to perform real-time analysis and prediction of the distribution network operating status based on pre-processed data; The cross-validation layer is used to dynamically modify and cross-validate the output parameters of the digital intelligence model library based on the adaptive kernel recursive least squares algorithm; The collaborative decision-making layer is used to combine cloud instructions with local key data to generate dynamic edge decisions and achieve cloud-edge collaboration; The cloud platform is used for full data storage, global collaborative scheduling, and visualization, and for two-way parameter interaction with the edge intelligence layer.

[0007] Preferably, the system also includes a dynamic adaptation mechanism for adjusting multi-level model parameters and cloud-edge collaboration strategies in real time according to the operating status of the distribution network, the status of equipment resources and the urgency of the task; the dynamic adaptation mechanism supports switching of multiple versions of lightweight models, including ultra-short-term, short-term and long-term load forecasting models, and sets trigger conditions based on time scales.

[0008] Preferably, the data processing platform includes a data acquisition layer, a data processing layer and a data transmission layer, wherein the data acquisition layer includes an event-triggered sampling mechanism, and its trigger function is: ; Where, For the current moment The monitoring parameters, is the sampling interval, is the standard deviation of historical data; The data processing layer includes anomaly detection methods and an outlier detection formula based on statistical features: ; Where, is the mean of historical data, Normal distribution of Quantile, Indicates the significance level.

[0009] Preferably, the digital intelligence model library is constructed based on partial least squares and improved temporal convolutional network, and the steps include: Let the input matrix and the corresponding output matrix , is the number of samples, is the number of input features for each sample, is the output dimension; Extracting latent variables from input and output data using partial least squares ; The latent variables extracted by partial least squares As the input of the improved time convolution network, the time convolution network is constructed and the hidden layer output vector is found to be and the output is Nonlinear relationship mapping, randomly given input weights and residuals ,choose is the activation function, and the number of neurons in the hidden layer is set to , calculate the output matrix of the hidden layer : ; Calculate the output weight matrix and the predicted output matrix : ; Where, for The Moore-Penrose generalized inverse matrix of , is the real output matrix; Finally, the PLS-ITCN data-driven model is obtained, and the model output is: ; Where, are the projection weights of the latent variables to the original output space, is the residual matrix.

[0010] Preferably, the implicit variable extraction formula is: ; Where, and are the implicit variables of the input matrix and output matrix respectively, and are the load matrices of the input matrix and output matrix respectively, are the projection weights of the original input features to the latent variables, are the projection weights of the latent variables to the original output space, is the number of implicit variables, and is the residual matrix.

[0011] Preferably, the cross-validation layer comprises: Online sparse modeling method: Initialize the time series data sequence and parameters; calculate the nearest neighbor distance and nearest neighbor vector of the new sample, and if it exceeds the expected range, it is determined to be an isolated point and discarded; dynamically update the online model parameters based on the prediction error and the minimum distance of the data dictionary; Online model parameter update formula: ; Where, For the The first iteration model parameter values, is the step size parameter, is the prediction error.

[0012] Preferably, the online sparse modeling method calculates the nearest neighbor distance and nearest neighbor vector of the new sample. If it exceeds the expected range, it is determined to be an isolated point and discarded. The specific range is defined as: ; ; Where, and is the quantization factor, and is the standard deviation, for The expected proximity distance at a given moment, for The expected nearest neighbor vector at time t.

[0013] Preferably, the collaborative decision-making layer includes a cloud-edge collaborative mechanism and a decision fusion strategy, wherein the cloud-edge collaborative mechanism includes SM2 signature verification of cloud instructions, triggering a local alarm and refusing execution when the verification fails, and the edge end uploads local data in a structured priority manner according to the cloud instructions or sends the cloud instructions to the edge online model, and dynamically adjusts the model parameters; the decision fusion strategy adopts a hybrid reasoning mode, and also has historical strategy caching and network disconnection decision logic.

[0014] Preferably, the devices of the perception control layer include existing devices and new devices, wherein the existing devices include smart meters, distribution terminals, power monitoring instruments and distribution network SCADA systems, etc., which are read through IEC61850 protocol, MQTT protocol or Modbus-TCP; the new devices are external environment monitoring devices, which are used to collect irradiance, wind speed and temperature / humidity of distributed photovoltaics. The installation method adopts spatial interpolation method to ensure that the distance from any location to the nearest monitoring point is less than the spatial threshold , calculate the minimum number of devices based on the Voronoi diagram covering method : ; in, 、 are the length and width of the power station, To round up.

[0015] Preferably, the feature extraction of the data processing station adopts an adaptive weighted algorithm and a kernel principal component analysis algorithm, and the weight calculation formula is: ; Where, is a set of sequences with the same characteristics, is the target variable, 、 、 They are The information entropy, correlation coefficient and volatility of is a hyperparameter that satisfies .

[0016] It can be seen from the above technical solutions that the present invention has the following beneficial effects:

[0017] This application provides a distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model. It collects multi-source data in real time through the perception and control layer, and the edge intelligence layer realizes local real-time analysis and cloud-edge collaborative decision-making. It combines the dynamic adaptation mechanism to optimize resource allocation, effectively solving the problems of insufficient real-time performance, concentrated computing pressure and global coordination difficulties of traditional systems, and has the advantages of improving response speed, reducing bandwidth costs and enhancing decision-making adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the implementation cases of the present invention or the technical solutions in the prior art, the following is a brief description of the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. Those skilled in the art can derive other drawings based on these drawings without inventive effort. Among them: Figure 1 A block diagram of a distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model provided by the present invention; Figure 2 A schematic diagram of the cloud-edge collaborative computing architecture of the present invention; Figure 3 A schematic diagram of the edge gateway computing architecture of the present invention; Figure 4 This is a flowchart of the digital intelligence model algorithm of the present invention; Figure 5 Schematic diagram of the online sparse modeling method of the present invention; Figure 6 This is a prediction diagram of user load and photovoltaic output based on the multi-level model of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1:

[0021] In order to solve the problems of insufficient real-time performance of traditional distribution networks, concentrated computing pressure, lack of edge device data transmission and storage security mechanism, as well as the existing centralized systems, edge independent systems, and cloud-edge collaborative architectures, such as poor real-time performance, lack of global coordination, and insufficient intelligent processing. Figure 1 As shown, the present invention proposes a distribution network edge gateway computing decision system based on a digital intelligence multi-level model, which includes: The perception control layer is used to collect perception data of the low-voltage distribution network, including operation data, equipment status data and external environment data; The edge intelligence layer includes a multi-level model connected in sequence. The multi-level model includes a data processing middle platform, a digital intelligence model library, a cross-validation layer, and a collaborative decision-making layer. Among them: the data processing middle platform is used to pre-process the perception data, including error processing, anomaly detection, data conversion and feature extraction, data fusion, and protocol encapsulation; the digital intelligence model library is used to perform real-time analysis and prediction of the distribution network operation status based on the pre-processed data; the cross-validation layer is used to dynamically modify and cross-validate the parameters of the output results of the digital intelligence model library based on the adaptive kernel recursive least squares algorithm; the collaborative decision-making layer is used to combine cloud instructions with local key data to generate dynamic edge decisions and realize cloud-edge collaboration; The cloud platform is used for full data storage, global collaborative scheduling, and visualization, and for two-way parameter interaction with the edge intelligence layer.

[0022] In addition, the system of the present invention also includes a dynamic adaptation mechanism, which is used to adjust the multi-level model parameters and cloud-edge collaboration strategies in real time according to the operating status of the distribution network, the status of equipment resources and the urgency of the task; the dynamic adaptation mechanism supports the switching of multiple versions of lightweight models, including ultra-short-term, short-term and long-term load forecasting models, and sets trigger conditions based on the time scale.

[0023] As can be seen from the above technical solution, the present invention proposes a distribution network edge gateway computing and decision-making system based on a multi-level digital intelligence model. The perception and control layer refers to a collection of sensor devices deployed at the end of the distribution network. It can be implemented using devices such as smart meters and environmental monitors. Its function is to build a global perception network to capture multidimensional operational information. The edge intelligence layer refers to a cluster of computing modules deployed on gateway devices. By building a multi-level model, it achieves data-to-decision conversion, solving the problem of a single processing flow in traditional architectures. The data processing center refers to a computing unit with data cleaning and feature extraction functions, using an event-triggered sampling mechanism to reduce redundant data transmission. The digital intelligence model library refers to a predictive model that integrates statistical methods and deep learning, improving prediction accuracy by combining partial least squares with a temporal convolutional network. The cross-validation layer refers to a parameter correction module based on an adaptive algorithm, eliminating abnormal data interference through online sparse modeling. The collaborative decision-making layer refers to a policy generation module that integrates cloud instructions and local data, using a hybrid inference model to ensure decision robustness and an SM2 signature verification mechanism to ensure instruction transmission security. The cloud platform refers to a resource scheduling center deployed on a remote server. The dynamic adaptation mechanism refers to switching the model's control module according to the urgency of the task, and achieving resource optimization configuration through lightweight model version management.

[0024] Compared with existing technologies, traditional centralized systems rely on cloud-based command transmission, resulting in decision-making delays. This solution achieves millisecond-level response by building a complete decision-making chain at the edge. Conventional independent edge systems lack model correction mechanisms, which can easily lead to accumulated prediction errors. This solution uses a cross-validation layer to establish a dynamic parameter adjustment mechanism to suppress the spread of bias. Existing cloud-edge collaborative architectures use fixed strategies that are difficult to adapt to environmental changes. This solution uses a dynamic adaptation mechanism to achieve real-time matching of model versions and resource status.

[0025] This application further proposes that the devices of the perception control layer include existing devices and new devices, among which the existing devices include smart meters, distribution terminals, power monitoring instruments and distribution network SCADA systems, which are read through the IEC61850 protocol, MQTT protocol or Modbus-TCP; the new devices are external environment monitoring devices used to collect irradiance, wind speed, temperature and humidity in the distributed photovoltaic area. The installation method adopts spatial interpolation method to ensure that the distance from any location to the nearest monitoring point is less than the spatial threshold δ, and the minimum number of devices is calculated based on the Voronoi diagram coverage method. , whose expression is , where L and W are the length and width of the power station respectively. To round up.

[0026] Spatial interpolation refers to the mathematical construction of a network of monitoring points in the spatial dimension. This can be achieved using Kriging interpolation or inverse distance weighted interpolation algorithms, and is used to generate optimal monitoring point locations when new equipment is deployed. The Voronoi diagram coverage method divides a planar area into several cells, each containing a monitoring point. Computational geometry algorithms can be used to generate polygonal coverage areas, which are used to determine the minimum number of devices that meet distance constraints. The IEC61850 protocol, MQTT protocol, or Modbus-TCP refer to power system communication standards, which can be used for substation automation, IoT data transmission, and industrial equipment communication, respectively, to enable data collection and integration of existing equipment.

[0027] Specifically, in the distributed photovoltaic scenario, first, based on the length L and width W of the power station area, combined with the set spatial threshold δ (for example, it can be set to 50 meters), the minimum number of monitoring devices required is calculated by the formula. New environmental monitoring equipment is deployed according to the coordinates determined by the spatial interpolation method. For example, in the photovoltaic panel array, the installation point is selected based on the distribution pattern of historical irradiance data. Existing equipment is connected to the system through standard communication protocols. For example, the distribution terminal uploads voltage data through the IEC61850 protocol, and the smart meter transmits power consumption information through the MQTT protocol. The monitoring area is divided by the Voronoi diagram to ensure that the distance between devices in each polygon area does not exceed value, thereby avoiding monitoring blind areas.

[0028] Through the above technical solution, this application effectively solves the problem of low deployment efficiency of environmental monitoring equipment in distributed photovoltaic scenarios. By using mathematical modeling methods to accurately calculate the installation location and quantity of equipment, it avoids monitoring blind spots and reduces hardware resource waste. The integration of standard communication protocols enables rapid access to multi-source heterogeneous equipment, providing a complete and reliable data foundation for subsequent data processing.

[0029] This application further proposes a data processing platform (first-level model) including a data collection layer, a data processing layer, and a data transmission layer: Data collection layer: The data collection layer communicates directly with the physical devices and comprehensively collects various parameters of the devices in the perception control layer. For the data of the external environment monitoring equipment, an event-triggered sampling mechanism is adopted, and the trigger function is .in, For the current moment The monitoring parameters, is the sampling interval, is the standard deviation of historical data. When the change in the monitoring parameter exceeds 3 times the standard deviation, timely data collection is carried out to reduce the amount of invalid data collection and improve data collection efficiency.

[0030] Among them, the event-triggered sampling mechanism refers to a technical means of dynamically adjusting the sampling frequency according to the change amplitude of the monitoring parameters. Specifically, it can be implemented by using the standard deviation threshold comparison method. When the change in the monitoring parameter exceeds three times the standard deviation, data collection is triggered, otherwise the sampling frequency is maintained at a low level.

[0031] Specifically, the data acquisition layer connects to devices such as smart meters via the IEC61850 or Modbus-TCP protocols, continuously acquiring raw monitoring data streams. The data processing layer dynamically processes the raw data streams: when the fluctuation range of parameters such as current and voltage is less than three times the historical standard deviation, the event-triggered sampling mechanism reduces the sampling frequency to one-third of the base frequency, for example, from 1Hz to 0.33Hz.

[0032] Data processing layer: The collected data has problems such as large data volume, redundancy and strong correlation. In the error processing stage, for gross errors, statistical analysis is used to identify and eliminate obvious abnormal data; for random errors, filtering algorithms are used for smoothing; for missing data, interpolation is performed based on data characteristics and historical patterns. Anomaly detection is based on the isolated point detection formula based on statistical characteristics. ,in is the mean of historical data, Normal distribution of Quantile, Indicates the significance level and accurately identifies anomalous data points. Data transformation uses normalization methods to unify data of different magnitudes and distributions into specific intervals for easy subsequent analysis and processing.

[0033] Among them, the anomaly detection method refers to the technical means of identifying data anomalies through statistical methods. Specifically, it can be implemented by adopting the isolated point judgment rule based on the confidence interval of the normal distribution, and the anomaly detection sensitivity is controlled by setting the significance level parameter α.

[0034] Specifically, when a temperature parameter change is detected that exceeds ±2.58 standard deviations of the historical mean, the anomaly detection method automatically marks the data as an outlier and triggers the data cleansing process. The processed and standardized data is uploaded to the edge intelligence layer via the MQTT protocol via the data transmission layer, while retaining a cache of the last 24 hours of data.

[0035] Feature extraction uses adaptive weighted algorithm and kernel principal component analysis algorithm, and the weight calculation formula is: ,in is a set of sequences with the same characteristics, is the target variable, 、 、 They are The information entropy, correlation coefficient and volatility of is a hyperparameter that satisfies , effectively extract key features of data.

[0036] Among them, the adaptive weighting algorithm refers to a calculation method that dynamically adjusts feature weights. Specifically, it can achieve multi-dimensional feature evaluation by measuring feature diversity through information entropy, assessing feature correlation through cross-correlation coefficients, and reflecting feature stability through volatility. The kernel principal component analysis algorithm refers to a nonlinear dimensionality reduction method. Specifically, it can achieve data noise reduction by transforming the original features into a high-dimensional space through kernel function mapping and then extracting the principal components. Information entropy is used to quantify the information richness of a feature. Specifically, the Shannon entropy formula can be used to calculate the uncertainty of the feature distribution. Cross-correlation coefficients are used to characterize the strength of the linear association between features and target variables. Specifically, the Pearson correlation coefficient can be used to calculate them. Volatility is used to measure the stability of feature data. Specifically, it can be achieved by calculating the ratio of the standard deviation of the feature value to the mean.

[0037] Specifically, during the feature extraction process for distribution network operation data, a feature sequence set is first constructed for the collected multi-dimensional data, such as current, voltage, and power. The information content of each feature is assessed by calculating its information entropy. For example, the information entropy of the current harmonic distortion rate feature is higher than that of the steady-state current feature. The correlation coefficient between each feature and the target variable is then calculated. For example, the correlation coefficient between the voltage sag feature and the equipment fault state is higher than that of the temperature feature. The volatility of each feature in the time dimension is also calculated. For example, the volatility of photovoltaic output data is significantly higher than that of distribution transformer load factor data. After weighted summing the three indicators according to preset hyperparameters, the feature weights are normalized to form a feature selection mechanism with dynamic adaptability. Finally, nonlinear dimensionality reduction is performed on the high-dimensional feature space through kernel principal component analysis. For example, a Gaussian kernel function is used to map the original features to a reproducing kernel Hilbert space and then extract the principal components.

[0038] Compared with existing technologies, traditional feature selection methods often use fixed weights or single-dimensional evaluation metrics, which are unable to dynamically adapt to the non-stationary characteristics of distribution network operation data. For example, conventional principal component analysis only considers linear dimensionality reduction methods that maximize variance, making it difficult to effectively handle nonlinear characteristics such as photovoltaic output and load fluctuations. Feature selection methods based on static thresholds tend to ignore the dynamic correlations between features, resulting in the loss of key features.

[0039] Through the above technical solution, this application effectively solves the problem of poor adaptability in feature selection for heterogeneous data in distribution networks. By integrating multi-dimensional evaluation indicators and a dynamic weighting mechanism, it can accurately identify key features in different operating scenarios. For example, it can automatically increase the weight of transient features in fault conditions and focus on energy efficiency feature analysis during steady-state operation. Combined with the dimensionality reduction processing of kernel principal component analysis, it can significantly improve edge computing efficiency while preserving the nonlinear characteristics of the data and providing reliable feature input for subsequent state prediction.

[0040] Data transmission layer: This layer aggregates processed data, removes duplication and redundancy, and encapsulates it according to specific protocols. Based on the multi-level model selected by the edge gateway, it efficiently transmits processed time-series data in real time, ensuring smooth data flow within the system.

[0041] This application further proposes a digital intelligence model library (secondary model) based on partial least squares (PLS) and improved temporal convolutional network (ITCN), which consists of data-driven models and artificial intelligence-driven models, including load forecasting models, photovoltaic output models, and other models. The steps include: Let the input matrix and the corresponding output matrix , is the number of samples, is the number of input features for each sample, is the output dimension; Extracting latent variables from input and output data using partial least squares ; The latent variables extracted by partial least squares As the input of the improved time convolution network, the time convolution network is constructed and the hidden layer output vector is found to be and the output is Nonlinear relationship mapping, randomly given input weights and residuals ,choose is the activation function, and the number of neurons in the hidden layer is set to , calculate the output matrix of the hidden layer : ; Calculate the output weight matrix and the predicted output matrix : ; Where, for The Moore-Penrose generalized inverse matrix of , is the real output matrix; Finally, the PLS-ITCN data-driven model is obtained, and the model output is: ; Where, are the projection weights of the latent variables to the original output space, is the residual matrix.

[0042] Among them, partial least squares is a statistical method that reduces the dimensionality of high-dimensional data by projection. Specifically, it can extract latent variables by calculating the covariance structure of the input matrix and the output matrix, which is used to eliminate multicollinearity and reduce data redundancy. The improved temporal convolutional network is a deep learning model with an expanded causal convolution structure. Specifically, it can be implemented using stacked convolutional layers and asymmetric residual connections to capture long-term dependencies in time series. Activation function This is a Sigmoid function that maps inputs to the 0-1 range through nonlinear transformations, enhancing the model's ability to fit nonlinear relationships. The number of hidden layer neurons, L, can be set based on computing resources or data size. For example, it can be set to an integer multiple of the number of input features to balance model complexity and generalization performance.

[0043] Specifically, the input and output matrices are decomposed into linear combinations of latent variables and residuals. The latent variables are extracted using partial least squares and then fed into the improved temporal convolutional network. Within the temporal convolutional network, the latent variables undergo multi-level convolution operations, and the output of each layer undergoes nonlinear transformations using activation functions, ultimately outputting predictions that match the dimensions of the latent variables. The output weight matrix is calculated using the generalized inverse matrix, allowing the model to approximate the true output with the least squares error. This model thus combines the linear dimensionality reduction advantages of partial least squares with the nonlinear time series modeling capabilities of the temporal convolutional network, forming a composite data-driven architecture.

[0044] Compared with existing technologies, traditional methods typically use partial least squares regression or a single time series model alone. The former has difficulty handling nonlinear relationships, while the latter lacks the ability to reduce the dimensionality of high-dimensional data. This solution, by cascading partial least squares with an improved temporal convolutional network, retains the advantages of linear feature extraction while enhancing the model's ability to capture complex temporal patterns. Existing technologies that use recurrent neural networks to process time series data suffer from vanishing gradients and low computational efficiency. This solution uses the dilated convolution structure of an improved temporal convolutional network to process time series data in parallel and reduce computational time.

[0045] Through the above technical solution, this application can effectively handle the coupling relationship between high-dimensional heterogeneous data and complex time series characteristics in distribution network operation, improving the accuracy of state prediction. At the same time, the model reduces the requirements for prior assumptions on data distribution through phased processing of linear and nonlinear modules, enhancing adaptability in different operating scenarios. In addition, the introduction of implicit variables reduces the interference of redundant features on model training, allowing edge gateways to achieve efficient modeling with limited computing resources.

[0046] This application further proposes a latent variable extraction formula, which is specifically expressed as the input matrix X and the output matrix Y are decomposed into a combination of latent variable matrices T and U and load matrices P and Q, and the residual matrices E and F are introduced to represent the decomposition error. The formula is as follows: .in, and are the implicit variables of the input matrix and output matrix respectively, and are the load matrices of the input matrix and output matrix respectively, are the projection weights of the original input features to the latent variables, are the projection weights of the latent variables to the original output space, is the number of implicit variables, and is the residual matrix.

[0047] Among them, latent variables refer to the low-dimensional core features extracted from the original data through mathematical decomposition. Specifically, this can be achieved using the partial least squares projection algorithm to eliminate redundant information and retain the correlation characteristics between input and output. The loading matrix refers to the projection coefficient matrix of the original data into the latent variable space. Specifically, it can be solved through an iterative optimization algorithm and is used to establish a mapping relationship between the original data and the latent variables. The residual matrix refers to the data residuals that are not explained by the model during the decomposition process. Specifically, it can be obtained through matrix subtraction operations. It is used to characterize the decomposition error and guide the correction of model parameters.

[0048] Specifically, in the distribution network edge gateway, the input matrix contains real-time monitoring data such as voltage and current, and the output matrix corresponds to the load forecast results. By synchronously decomposing the input and output data into a linear combination of latent variables and the load matrix, the potential correlation between the two can be effectively captured. For example, in scenarios where distributed photovoltaic output fluctuates, latent variables can extract the nonlinear coupling characteristics of light intensity and power output. The residual matrix is used to quantify model fitting deviations. When the residual exceeds a preset threshold, the model's dynamic correction mechanism is triggered. For example, when thunderstorms cause data anomalies, a sudden increase in residuals can be identified as an external interference event.

[0049] Compared with existing technologies, traditional principal component analysis only reduces the dimensionality of input data without considering the correlation of output variables, causing feature extraction to deviate from the actual prediction target. This solution jointly decomposes the input and output matrices, allowing implicit variables to simultaneously reflect the coupling relationship between device status and operational objectives. For example, in a voltage sag event, implicit variables can simultaneously characterize fault characteristics and protection action requirements. Furthermore, existing technologies often ignore the impact of residuals on model robustness. This solution uses the residual matrix to monitor model errors in real time. For example, when communication interruptions lead to data loss, residual changes can trigger local caching strategies.

[0050] Through the above technical solutions, this application can solve the problem of low model efficiency caused by high-dimensional edge data and limited computing resources. For example, in the task of substation topology analysis, implicit variables compress the original 128-dimensional data to 8 dimensions, significantly reducing computational complexity. At the same time, the dynamic monitoring mechanism of the residual matrix can improve the model's adaptability to abnormal data. For example, when equipment failure causes data jumps, residual analysis can avoid misjudging it as normal load fluctuations, thereby improving decision reliability.

[0051] This application further proposes that the cross-validation layer (three-level model) adopts an online sparse modeling method including: Initialize time series data sequence and parameters: Initialize time series data sequence And step size parameter η, kernel width σ, quantization factor threshold θ, quantization factor , Parameters such as initial data dictionary , the coefficient of the element in the data dictionary .

[0052] calculate Time by The standard deviation of the constructed series ;calculate Time by The standard deviation of the constructed series .

[0053] Calculate the nearest neighbor distance of the new sample and neighbor vectors , and its corresponding range: and If both are outside the above expected range, then the new data sample It is an isolated point and will be discarded by the iterative learning process and will not be used for parameter update.

[0054] The nearest neighbor distance range refers to the data acceptance interval set by the dynamic standard deviation and quantization factor. This can be achieved by using a sliding window statistical method to update the standard deviation parameter in real time, eliminating the lack of adaptability caused by fixed thresholds. The quantization factor refers to the adjustment coefficient that controls the boundaries of the range. It can be determined through offline training or online adaptive algorithms to balance the strictness of data screening and the generalization ability of the model. The expected nearest neighbor vector refers to a reference vector sequence generated based on the distribution of historical data. It can be generated using a time series prediction model and is used to characterize data association characteristics under normal operating conditions.

[0055] Specifically, when processing the edge data stream of the distribution network in real time, this method first initializes the time series data sequence and model parameters. For each newly arrived sample, its Euclidean distance with the samples in the current data dictionary is calculated to obtain the nearest neighbor distance statistics, and the nearest neighbor vector direction features are extracted at the same time. When the nearest neighbor distance or vector direction of a new sample exceeds the dynamic range, it is determined to be an outlier and directly discarded to prevent abnormal data from entering the model update process. For data that meets the range, the prediction error is further calculated, and the model parameters are dynamically adjusted according to the error gradient. By continuously updating the standard deviation and expected value, it can adapt to changes in the operating state of the distribution network and ensure the robustness of the sparse modeling process.

[0056] Compared with existing technologies, traditional outlier detection methods typically use fixed thresholds or static statistical models, which are prone to misjudgments or missed detections in dynamic distribution network scenarios, resulting in reduced model accuracy. This solution introduces a dynamic range calculation mechanism based on time series, enabling the anomaly determination boundary to automatically adjust as the data distribution changes, effectively addressing the model's lack of adaptability due to environmental disturbances.

[0057] Through the above technical solution, this application can achieve real-time filtering of abnormal data and online optimization of model parameters under conditions of limited edge computing resources. For example, in scenarios where distributed photovoltaic output suddenly changes or load fluctuates, abnormal data generated by sensor failures or communication interference can be quickly identified to prevent interference with load forecasting models. This also reduces the consumption of ineffective computing resources and improves the decision-making efficiency and reliability of edge gateways.

[0058] Calculate data samples The prediction error ,calculate Minimum distance to the current data dictionary Dynamically update online model parameters if Keep the current data dictionary unchanged; otherwise update the current data dictionary to: , the update formula is ,in For the The first iteration model parameter values, is the step size parameter, is the prediction error.

[0059] Among them, the online sparse modeling method refers to the active removal of abnormal samples through the nearest neighbor vector screening mechanism during the time series data processing process. Specifically, it can be implemented by using a sliding time window combined with a kernel density estimation method. This technology can solve the problem of model overfitting caused by data redundancy. The online model parameter update formula refers to a gradient descent algorithm based on prediction error feedback. Specifically, it can be implemented by an adaptive learning rate adjustment strategy. This technology can dynamically adjust the parameter update step size according to the real-time error to avoid oscillation divergence during the model iteration process. The nearest neighbor distance refers to the Euclidean distance or cosine similarity of the sample vector in the feature space. Specifically, it can be implemented by the k-nearest neighbor algorithm. This technology can effectively identify abnormal data points. The expected range refers to a dynamic confidence interval constructed based on historical statistics. Specifically, it can be implemented by using the moving average method and standard deviation calculation. This technology can adapt to the characteristics of data distribution changing over time.

[0060] Specifically, the cross-validation layer uses online sparsification modeling to cleanse the time series data stream in real time. For example, when newly collected voltage fluctuation data exceeds a preset threshold, it is identified as noise data and automatically filtered out. The online model parameter update formula continuously optimizes model parameters through an error feedback mechanism. For example, when the load forecast error exceeds a set threshold, the step size parameter is automatically increased to accelerate convergence. This technical solution, through online data screening and dynamic parameter adjustment, effectively addresses the problem of reduced prediction accuracy in traditional edge computing models caused by fluctuating data quality.

[0061] This application further proposes a collaborative decision-making layer (four-level model) including a cloud-edge collaborative mechanism and a decision fusion strategy. The cloud-edge collaborative mechanism includes SM2 signature verification of cloud instructions. When the verification fails, a local alarm is triggered and execution is refused. The edge end uploads local data in a structured manner according to priority based on cloud instructions or sends cloud instructions to the edge online model, and dynamically adjusts model parameters. The decision fusion strategy adopts a hybrid reasoning mode, and also has historical strategy caching and network disconnection decision logic.

[0062] SM2 signature verification refers to the process of digitally verifying the signature of cloud-based commands based on the national secret SM2 algorithm. This can be implemented using elliptic curve cryptography to ensure the legitimacy and integrity of the command source and prevent malicious command injection. Local alarm triggering automatically activates the alarm module when signature verification fails. This can be achieved by combining preset alarm thresholds with audio and visual prompts to prevent illegal command execution from causing system anomalies. Priority structured upload classifies local data for transmission based on its criticality and timeliness. This can be achieved by integrating data tags with bandwidth allocation strategies to optimize data transmission efficiency and reduce bandwidth usage. Dynamic model parameter adjustment updates the weights or structure of edge models in real time based on cloud-based commands. This can be achieved using online learning algorithms to adapt to changes in power grid operating conditions. Hybrid inference mode refers to a comprehensive decision-making approach that combines real-time data and historical strategies. This can be achieved by integrating a rules engine with a machine learning model to improve decision robustness. Historical strategy caching stores historically optimal decision plans as a backup. This can be implemented using a time-series database to ensure that backup plans can be used in the event of real-time analysis failure. The network disconnection decision logic refers to enabling the local autonomous decision-making process when the network is interrupted. Specifically, it can be implemented by coordinating preset emergency rules with offline models to maintain the basic functions of the system.

[0063] Specifically, when the cloud sends a control command to the edge, the command's digital signature is first verified using the SM2 algorithm. If the signature verification fails, a local alarm is triggered and the command is rejected, preventing unauthorized operation. After successful verification, the edge uploads local data according to the command's pre-set priorities. For example, device fault alarm data is prioritized while routine monitoring data is delayed, optimizing network resource allocation. Simultaneously, parameter adjustment commands issued by the cloud are synchronized with the edge's online model, for example, dynamically modifying the weight coefficients of the load forecasting model to adapt to fluctuations in PV output. During the decision-making process, a hybrid inference model combines real-time analysis results with historical caching strategies. For example, decisions are generated by combining the current grid load status with historical optimal solutions from similar scenarios. If the network connection is lost, the disconnection decision logic immediately activates locally cached historical strategies and pre-set rules, for example, maintaining basic control functions based on the last valid cloud command and local data.

[0064] Compared with existing technologies, conventional cloud-edge collaborative architectures lack effective security verification mechanisms and are unable to identify illegal instructions. However, this solution strengthens the credibility of the source of instructions through SM2 signature verification technology. The data upload strategy of existing technologies usually adopts a fixed transmission mode, which can easily cause bandwidth waste or delay of critical data. This solution significantly improves data transmission efficiency through priority classification upload. In addition, existing edge systems are prone to functional paralysis when the network is interrupted. The network disconnection decision logic of this solution achieves downgraded operation through local caching strategies and emergency rules to ensure basic system availability.

[0065] This application further proposes that the cloud platform be responsible for full data storage, global collaborative scheduling, and visual display. By establishing an efficient data storage architecture, massive power data is classified and stored to facilitate rapid retrieval and analysis. Artificial intelligence algorithms are used to achieve global collaborative scheduling and optimize power resource allocation. Visualization technology is used to present the operating status of the power system to users in the form of intuitive charts and graphs to assist in decision-making. At the same time, the cloud platform and the edge intelligence layer conduct two-way parameter interaction to achieve continuous optimization of the model and improve system performance.

[0066] Taking the distributed photovoltaic and user load model on June 1, 2022 as an example, user load data is collected using smart meters within the substation area, and prediction results are output every 15 minutes. Distributed photovoltaic data, including historical data and real-time environmental data, is also collected and output every 15 minutes. During operation, the perception and control layer continuously collects data and transmits it to the edge intelligence layer. The data processing center preprocesses the data, the digital intelligence model library predicts load and photovoltaic output, and the cross-validation layer corrects the prediction results. The collaborative decision-making layer combines cloud instructions and local data to make decisions, and the cloud platform implements data storage, scheduling, and visualization. Judging from the experimental results (such as the output result diagram of user load and photovoltaic output), the third-level model has a corrective effect on the output results of the second-level model, effectively improving the accuracy of the model prediction results and verifying the effectiveness and reliability of the system.

[0067] Through the above technical solution, this application effectively solves the problem of lack of cloud command security verification, avoids system risks caused by illegal operations; optimizes data transmission efficiency and reduces network bandwidth pressure; enhances the edge's autonomous decision-making ability under network anomalies, and ensures the continuity and stability of distribution network operation.

[0068] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0070] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0071] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.

Claims

1. A distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model, characterized in that: include: The perception control layer is used to collect perception data of the low-voltage distribution network, including operation data, equipment status data and external environment data; The edge intelligence layer includes a multi-level model connected in sequence. The multi-level model includes a data processing platform, a digital intelligence model library, a cross-validation layer, and a collaborative decision-making layer, where: The data processing platform is used to pre-process the perception data, including error handling, anomaly detection, data conversion and feature extraction, data fusion and protocol encapsulation; A digital intelligence model library, used to perform real-time analysis and prediction of the distribution network operating status based on pre-processed data; The cross-validation layer is used to dynamically modify and cross-validate the output parameters of the digital intelligence model library based on the adaptive kernel recursive least squares algorithm; The collaborative decision-making layer is used to combine cloud instructions with local key data to generate dynamic edge decisions and achieve cloud-edge collaboration; The cloud platform is used for full data storage, global collaborative scheduling, and visualization, and for two-way parameter interaction with the edge intelligence layer.

2. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The system also includes a dynamic adaptation mechanism for adjusting multi-level model parameters and cloud-edge collaboration strategies in real time according to the operating status of the distribution network, the status of equipment resources, and the urgency of the task; the dynamic adaptation mechanism supports switching of multiple versions of lightweight models, including ultra-short-term, short-term, and long-term load forecasting models, and sets trigger conditions based on time scales.

3. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The data processing platform includes a data acquisition layer, a data processing layer, and a data transmission layer. The data acquisition layer includes an event-triggered sampling mechanism, and its trigger function is: ； Where, For the current moment The monitoring parameters, is the sampling interval, is the standard deviation of historical data; The data processing layer includes anomaly detection methods and an outlier detection formula based on statistical features: ； Where, is the mean of historical data, Normal distribution of Quantile, Indicates the significance level.

4. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The digital intelligence model library is constructed based on partial least squares and improved temporal convolutional network, and its steps include: Let the input matrix and the corresponding output matrix , is the number of samples, is the number of input features for each sample, is the output dimension; Extracting latent variables from input and output data using partial least squares ; The latent variables extracted by partial least squares As the input of the improved time convolution network, the time convolution network is constructed and the hidden layer output vector is found to be and the output is Nonlinear relationship mapping, randomly given input weights and residuals ,choose is the activation function, and the number of neurons in the hidden layer is set to , calculate the output matrix of the hidden layer : ； Calculate the output weight matrix and the predicted output matrix : ； Where, for The Moore-Penrose generalized inverse matrix of , is the real output matrix; Finally, the PLS-ITCN data-driven model is obtained, and the model output is: ； Where, are the projection weights of the latent variables to the original output space, is the residual matrix.

5. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 4 is characterized in that: The implicit variable extraction formula is: ； Where, and are the implicit variables of the input matrix and output matrix respectively, and are the load matrices of the input matrix and output matrix respectively, are the projection weights of the original input features to the latent variables, are the projection weights of the latent variables to the original output space, is the number of implicit variables, and is the residual matrix.

6. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The cross validation layer includes: Online sparse modeling method: Initialize the time series data sequence and parameters; calculate the nearest neighbor distance and nearest neighbor vector of the new sample, and if it exceeds the expected range, it is determined to be an isolated point and discarded; dynamically update the online model parameters based on the prediction error and the minimum distance of the data dictionary; Online model parameter update formula: ； Where, For the The first iteration model parameter values, is the step size parameter, is the prediction error.

7. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 6 is characterized in that: The online sparse modeling method calculates the nearest neighbor distance and nearest neighbor vector of the new sample. If it exceeds the expected range, it is determined to be an isolated point and discarded. The specific range is defined as: ； ； Where, and is the quantization factor, and is the standard deviation, for The expected proximity distance at a given moment, for The expected nearest neighbor vector at time t.

8. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The collaborative decision-making layer includes a cloud-edge collaborative mechanism and a decision fusion strategy, wherein the cloud-edge collaborative mechanism includes SM2 signature verification of cloud instructions, triggering a local alarm and refusing execution when the verification fails. The edge end uploads local data in a structured manner according to priority based on cloud instructions or sends cloud instructions to the edge online model, and dynamically adjusts model parameters; the decision fusion strategy adopts a hybrid reasoning mode, and also has historical strategy caching and network disconnection decision logic.

9. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The devices in the perception control layer include existing devices and new devices. The existing devices include smart meters, distribution terminals, power monitoring instruments and distribution network SCADA systems, which are read through IEC61850 protocol, MQTT protocol or Modbus-TCP; the new devices are external environment monitoring devices used to collect irradiance, wind speed and temperature / humidity in the distributed photovoltaic area. The installation method adopts spatial interpolation method to ensure that the distance from any location to the nearest monitoring point is less than the spatial threshold. , calculate the minimum number of devices based on the Voronoi diagram covering method : ； in, 、 are the length and width of the power station, To round up.

10. The distribution network edge gateway computing and decision-making system based on a digital intelligence multi-level model according to claim 1 is characterized in that: The feature extraction of the data processing station adopts the adaptive weighted algorithm and the kernel principal component analysis algorithm, and the weight calculation formula is: ； Where, is a set of sequences with the same characteristics, is the target variable, 、 、 They are The information entropy, correlation coefficient and volatility of is a hyperparameter that satisfies .

Images